The Taxonomy That Ate the Problem

O'Neil's central gift was a naming convention. Opacity, Scale, Damage—three criteria, capitalized and repeated until they work like a diagnostic checklist. A WMD is any model that scores high on all three. The framework was elegant enough that it escaped the book almost immediately, entering policy conversations, EU regulatory language, and the vocabulary of a generation of data-ethics researchers who needed something sturdier than "algorithmic harm" and less loaded than "digital redlining." The taxonomy did real work. It let a journalist or a city council member point at a scheduling algorithm or a recidivism score and say, precisely, *why* it was dangerous rather than merely *that* it was. Ten years on, the framework still circulates—but the world it was built to describe has mutated in ways that strain it.

Consider what the book saw clearly. O'Neil's chapter on teacher evaluation remains a small masterpiece of institutional reporting. Tim Clifford, a veteran New York English teacher, scored a 6 out of 100 one year and a 96 the next, teaching identically both times. The anecdote doesn't need updating; it needs no temporal pivot to land. The same is true of her analysis of PredPol and predictive policing, where she identified the feedback loop—historical arrest data trains the model, the model sends cops back to the same blocks, new arrests confirm the model—with a clarity that ProPublica's later COMPAS investigation would corroborate in detail. She wrote that "data scientists are stitching this status quo of the social order into models…that hold ever-greater sway over our lives," and she was describing not a future risk but an already-operational machine. Her chapter on for-profit college advertising, too, pinpointed the precise mechanism: machine-learning systems that "probe for deeper patterns," gauge "weaknesses and vulnerabilities," and pursue "the most efficient path to exploit them." The predatory-ad pipeline she mapped has only grown more sophisticated, not less.

What the book could not have built its taxonomy around, because it did not yet exist at consumer scale, is generative AI. O'Neil's WMDs are *decision* systems—models that sort, score, and gate. They say yes or no: you get the loan, you don't; you're high-risk, you're not. The systems dominating public anxiety in 2026 are *production* systems—models that generate text, images, code, legal briefs, synthetic voices. They don't slam doors; they flood rooms. The Opacity criterion still applies, arguably more than ever, since even the engineers training large language models cannot fully explain their outputs. But Scale and Damage work differently when the model isn't sorting humans into buckets but producing artifacts that reshape labor markets, intellectual property, and the information environment wholesale. O'Neil's framework assumed a world where the algorithm was a gatekeeper. It has less to say about a world where the algorithm is also a ghostwriter, a forger, and an ersatz colleague. This is not a failure of imagination so much as a consequence of writing about the present tense: the book's strength was its refusal to speculate, its insistence on documented harm, and that discipline necessarily bounded its horizon.

There is a subtler gap, too. O'Neil's proposed remedy—algorithmic auditing, burden-of-proof regulation, stakeholder input—was framed as a conversation among engineers, ethicists, and affected communities. "We will probably never have a simple and universally agreed upon definition of what makes an algorithm fair," she wrote. "But thank goodness we're finally having the conversation." The conversation did happen. The EU passed the AI Act. The White House issued its Blueprint for an AI Bill of Rights. Algorithmic auditing firms sprouted. And yet the WMDs she catalogued—credit scoring proxies, scheduling software that treats workers as inventory, recidivism models encoding racial disparity—remain largely operational, their owners having absorbed the new vocabulary of "responsible AI" without materially changing the models. The book assumed that transparency would generate accountability, that sunlight was a sufficient disinfectant. What the decade has demonstrated is that sunlight can also just become ambient lighting—something everyone adjusts to, squinting a little, while the machinery hums on unchanged.

The book sits in the corpus alongside Shoshana Zuboff's *The Age of Surveillance Capitalism* and James Bridle's *New Dark Age*, but it occupies a different register: less theoretical, less panoramic, more prosecutorial. Zuboff wanted to name an entire economic logic; Bridle wanted to describe an epistemological crisis; O'Neil wanted to show you the spreadsheet that got a teacher fired. That specificity is why the book still gets assigned in data-science programs when the grander treatises gather dust. It is also why its limits matter. The real question the book deposits on the shelf, one it was too early and too honest to answer, is whether the problem was ever the math at all—or whether the math was always just a faster, cheaper way of executing decisions that institutions were determined to make regardless.